Electronics and computer hardware

Each stage has a check. Do not move on until it passes.

Mains and battery hazards are on electricity. These are the ones particular to electronics and chip-making.

A device can build a computer if it has four properties:

• Gain: one output can drive several inputs, so a signal can branch.
• A threshold: an input a little off its ideal value still gives a clean output, so noise does not build up over thousands of stages.
• One-way action: the output does not disturb the input.
• Inversion: some arrangement gives NOT.

A relay has all four through the gap between its pull-in and drop-out voltages. A valve and a transistor have them through their steep switching curves. The table compares the switches that matter. How valves are made is on electronics.

Parts: a coil on an iron core, an armature the coil pulls in against a spring, and contact springs the armature moves. A normally open contact closes when the coil is energised, a normally closed one opens, and a changeover contact does both. A relay pulls in above one voltage and drops out only below a much lower one: a common 12 V relay (Omron G5LE) must pull in by 9 V and must release by 1.2 V. That gap cleans up a noisy signal.

• AND: contacts in series. OR: contacts in parallel. NOT: a normally closed contact.
• Memory: a relay with one of its own normally open contacts wired across its set switch holds itself in until a reset contact breaks the circuit. It keeps one bit for as long as the power stays on.
• Adding: route the carry through changeover contacts that the two input bits have already set. The carry then runs the length of the adder through closed contacts in one relay operating time, instead of one per bit.
• Scale: Harry Porter's 8-bit relay computer has 415 identical four-pole 12 V relays, 8-bit data and 16-bit addresses, and draws about 160 W (documentation).

Rules that keep it working:

• A diode across every DC coil, cathode to the positive side. When the coil switches off, its collapsing field makes a spike of hundreds of volts that otherwise arcs and burns the contact driving it.
• Silver contacts need a minimum current, about 100 mA on many relays, to keep their surface clean. At logic currents of a few milliamps, use gold-plated or reed contacts.
• A contact rated for 250 V AC may be rated for only 30 V DC at the same current, because a DC arc does not go out by itself.
• Time each step for operate time plus bounce: about 20 ms per step for general-purpose relays.

Silicon has 5 × 10²² atoms per cm³. Pure silicon at room temperature has only about 10¹⁰ free charge carriers per cm³ and barely conducts.

• n-type: replace a few atoms with phosphorus or arsenic, which have five outer electrons; each gives a free electron.
• p-type: replace them with boron, which has three; each leaves a hole that moves like a positive charge.
• Working doping runs from 10¹⁵ to 10²⁰ atoms per cm³: one dopant in 50 million silicon atoms, up to one in 500. An unplanned impurity at 20 parts per billion already equals the lightest doping. That is why purification decides everything (semiconductors).

Where p meets n, carriers cross over and leave a thin depleted layer with a built-in voltage. Under forward bias (p positive), current flows above about 0.6–0.7 V in silicon and 0.2–0.3 V in germanium. Under reverse bias, almost nothing flows until breakdown. At room temperature, each extra 60 mV of forward voltage multiplies the current by ten. At a fixed current the forward voltage falls about 2 mV per °C, which gives electronic thermometers and causes thermal runaway in transistors.

Natural crystals rectify with no purification. A fine wire pressed on galena (lead sulfide) or iron pyrite detects radio. This cat's-whisker detector was the first semiconductor device (radio).

A diode is a single junction. Exceeding its maximum current or its reverse voltage is the usual way it dies, so record both with every salvaged diode.

Bench check with a meter's diode range: a good silicon diode reads about 0.5–0.7 V one way and open the other. A shorted one reads near zero both ways; a burnt-out one reads open both ways.

Bipolar transistor (BJT). A thin base layer sits between emitter and collector. A small base current controls a collector current β times larger: 100 to 300 for a small-signal part such as the 2N3904 at 10 mA. When on, the base–emitter junction sits at about 0.65 V, and a saturated switch drops about 0.1–0.3 V from collector to emitter. That 0.65 V falls 2 mV per °C, so a hot transistor draws more current, heats further and can run away. Put a resistor in the emitter and size the heatsink.

MOSFET. A gate electrode sits on a layer of oxide over the silicon between two doped regions. A gate voltage above the threshold pulls carriers into a thin channel under the oxide and connects source to drain. No steady current flows into the gate. That is why MOSFETs became the switch in every chip, and why they die of static: the oxide breaks down at tens of volts on a discrete part and at only a few volts on a modern chip. A power MOSFET switches tens of amps with milliohms of on-resistance. A logic-level MOSFET is fully on with 4.5 V or less on its gate.

The transistor a laboratory can make by hand is the germanium alloy-junction type. Beads of indium on both faces of a thin slice of n-type germanium are heated to around 500 °C, and they alloy into it to form the p-type emitter and collector.

Sodium wrecks MOS devices. A trace from skin or impure chemicals drifts through the oxide and shifts the threshold. This made 1960s MOS chips unreliable until clean processing cured it.

CMOS pairs a p-channel and an n-channel MOSFET so that in either steady state one of them is off. A CMOS gate draws current only while it switches. That is why a chip of billions of transistors can idle on milliwatts, and why CMOS, invented in 1963, replaced every other logic form.

• NAND takes 4 transistors and builds everything. NOT is a NAND with its inputs joined. AND is NAND followed by NOT. OR is a NAND of the two inputs inverted.
• Counts that size a design: an inverter is 2 transistors, a NAND or NOR 4, a full adder 28 in the standard mirror design, a static RAM bit 6, and a D flip-flop about 20.
• Each time a node charges and discharges, it uses energy C × V². Power is about activity × C × V² × frequency, so halving the voltage quarters the energy.

Why clock rates stopped rising. From the 1970s, each shrink let the voltage fall with the size, so power per unit area stayed constant (Dennard scaling, 1974). But a transistor cannot turn off more sharply than about 60 mV per tenfold change in current at room temperature. So supply voltages stalled near 0.7–1 V in the mid-2000s, and clock rates stalled at 3–5 GHz. Gains since then come from more processor cores, specialised circuits and new transistor shapes (Since 2000).

Most salvaged logic chips are the 7400 family (numbered 74xx00 onwards) or the older 4000 family. Families can be mixed only where one's output voltages meet the other's input thresholds. The table gives them for a 5 V supply unless it says otherwise.

Wiring rules, each one a common fault:

1. Leave no input floating. An open CMOS input drifts to mid-supply, where both transistors conduct: the chip heats, draws current and oscillates. Tie every unused input to the supply or to ground through 1–10 kΩ (TI SCBA004).
2. Put a 100 nF ceramic capacitor across the supply pins of every chip, as close as the leads allow, and 10–100 µF where power enters the board.
3. Debounce every switch. A mechanical contact bounces for milliseconds and is counted many times.
4. Pass every input not timed by the clock through two flip-flops in series before using it. Otherwise a change that lands on a clock edge can leave a flip-flop undecided (metastable).
5. Keep to the timing budget. The clock period must exceed the flip-flop's clock-to-output delay, plus the slowest logic path, plus the setup time. Measure the slowest path; do not guess it.

Choose memory by what it must survive. Working memory must be fast; the archive must outlast its machines. Core memory and paper are the two kinds a small society can make. Everything else is salvaged, and copied before it fails.

• Flash loses its data when left unpowered. A worn consumer SSD need only hold its data for 1 year unpowered at 30 °C, and an enterprise SSD for 3 months (JEDEC JESD218). Retention roughly halves for every 5 °C warmer. Power up salvaged drives and copy them off now.
• Microcontroller flash lasts much longer. The ATmega328P in the Arduino Uno is rated to keep its program 20 years at 85 °C and 100 years at 25 °C, but for only 10,000 rewrites.
• DRAM forgets in milliseconds. The memory controller refreshes every row at least every 64 ms.
• Test every memory before trusting it. Write and read back all zeros, all ones, alternating patterns, and each location's own address, which catches faults in the address lines.

The parts:

• a program counter, holding the address of the next instruction;
• an instruction register and decoder;
• a few registers;
• an arithmetic-logic unit: an adder plus AND, OR, XOR and shift;
• memory address and data paths;
• control: a state machine that steps through fetch, decode and execute, built from gates or from a small memory holding the steps (microcode);
• a clock, a reset that starts the counter at a known address, and input/output.

The instruction set can be small. Load, store, add, subtract, AND, OR, XOR, shift, jump, branch if zero or negative, and input/output are enough for any program, given enough memory.

Build to an open, existing instruction set rather than inventing one. RISC-V's base set, RV32I, has 40 instructions (RISC-V). It was ratified in 2019 and anyone may implement it free. The GCC and LLVM compilers and existing operating systems already target it, so a home-built RISC-V processor runs software nobody has to rewrite.

The first input/output is a serial line: one wire each way. Each character goes as a start bit, 8 data bits and a stop bit, so 9,600 bits per second carries 960 characters a second. It comes straight from the telegraph and teleprinter, needs one shift register at each end, and every salvaged microcontroller and terminal speaks it.

Bootstrapping is a ladder of programs, each written with the one before:

1. Toggle switches enter a short loader directly into memory, one word at a time.
2. The loader reads in a monitor: tens to hundreds of bytes that examine and change memory and accept hexadecimal from the serial line.
3. A hex loader, then an assembler, then a compiler follow, each assembled or compiled by the stage before.

Keep the whole chain as printed source with checksums. In 2023 GNU Guix rebuilt its whole software distribution from source, starting from a 357-byte program small enough to check by hand (Guix). Keep to that standard: a seed a person can verify, and everything above it rebuilt from source a person can read.

A microcontroller is a complete computer on one chip: processor, program memory, RAM, timers and serial lines. Nearly every appliance made since the 1990s has at least one. The Arduino Uno's ATmega328P has 32 KB of program flash and 2 KB of RAM, and runs at 16 MHz. With a programmer, salvaged microcontrollers run new programs. Keep their datasheets on paper: without its pinout and instruction set, a chip cannot be used.

Order of salvage: first copy the data off drives and memory cards, because it is fading. Then take boards with processors and memory, then power supplies, transistors, relays and passive parts. Store them as on semiconductors.

Soldering. Tin-lead 63/37 solder melts at a single temperature, 183 °C, so joints freeze cleanly. Lead-free SAC305 melts at 217–220 °C and needs a hotter iron. Run the tip at about 320–370 °C, heat pad and lead together for 2–3 seconds, and feed solder into the joint, not onto the iron. A good joint is a smooth, concave fillet that wets both pad and lead. A grey, grainy or ball-shaped joint is cold and will fail.

Etching your own boards needs copper-clad board, a resist pattern (printed toner or photoresist) and an etchant. Use ferric chloride, or cupric chloride, which air and hydrochloric acid regenerate so it can be used indefinitely. That suits a settlement with no supply of fresh chemicals.

Tools, in the order you need them:

• a multimeter;
• a logic probe, which you can build from two LEDs and a comparator;
• an oscilloscope with a bandwidth of at least 5 times the clock frequency. Showing a rise time tᵣ needs a bandwidth of about 0.35 ÷ tᵣ;
• for buses, a logic analyser, which a salvaged microcontroller can be programmed to be.

Every stage must succeed on every wafer, and a chip is only as good as its worst stage. The purification chemistry and its hazards are on semiconductors. Process recipes, etch rates and oxide-growth calculators are in the BYU cleanroom pages under Further.

A particle bigger than a feature kills the circuit under it. Clean rooms are graded by particles per cubic metre of air. ISO class 5 allows 3,520 particles of 0.5 µm or larger; ordinary room air is around class 9, at 35,200,000: ten thousand times dirtier (ISO 14644-1). Modern fabs keep wafers in sealed carriers and enclosures cleaner still.

• Filtered air flows downward from HEPA filters across the whole ceiling, at slightly higher pressure than outside, so leaks blow outward.
• People are the main source of particles: wear full suits, hoods, masks and boots, and enter through a gowning room or air shower.
• Rinse with ultrapure water at 18.2 MΩ·cm, the resistivity of water with nothing dissolved in it.
• Use yellow light where photoresist is handled, because blue and ultraviolet light expose it.
• Keep bare skin away from wafers, both for particles and for sodium, which ruins MOS transistors.

What a laboratory can reach without a full clean room: a laminar-flow hood over the work gives class-5 air where it matters, and features of 5–10 µm tolerate the dust that remains. That is the scale of the 1971 Intel 4004, and of a 2021 home-garage chip with 100 transistors made on a 10 µm self-aligned polysilicon-gate process (Zeloof).

The smallest half-pitch one exposure can print is about k₁ × λ ÷ NA. Here λ is the wavelength, NA is the lens's numerical aperture, and k₁ is a process factor that cannot go below 0.25 and is about 0.3 in production. Shorter light and wider lenses are the whole history of chip scaling. Below the single-exposure limit, features are split across two or four exposures (multiple patterning), at rising cost.

Node names stopped measuring anything in the mid-1990s. A "5 nm" process has a contacted gate pitch of 48 nm and a tightest metal pitch of 36 nm (IEEE Spectrum).

Making masks without a mask shop: draw the pattern large and reduce it photographically through a good lens, as early chip masks were made. Contact printing, with the mask pressed against the resist and ultraviolet light from a mercury lamp, reaches about 1–2 µm.

Each line is from its primary source.